Low-Latency 64-Parallel 4096-Point Memory-Based FFT for 6G

Abstract

This paper presents a novel 64-parallel 4096-point radix-2 memory-based fast Fourier transform (FFT) architecture for 6G. This approach is the first one to use 64 parallel branches in memory-based architectures. The challenge of designing a memory-based FFT with such a high parallelization has been accomplished by paying special attention to the large number of memories in parallel. Their control has been simplified by using the same read and write address for all of them thanks to the perfect shuffle permutation, and they are organized in groups to eliminate unnecessary registers. Likewise, a novel design for the rotation memories allows for reusing rotation coefficients among parallel rotators, and a new design for the circular counter that controls the architecture is presented. The proposed FFT architecture has been implemented on a Virtex 7 field-programmable gate array (FPGA). Experimental results reveal that the proposed architecture achieves the lowest latency in clock cycles and the highest throughput in samples per clock cycle among memory-based FFT architectures so far.